IEEJ Journal of Industry Applications

Vol.12 No.4 pp.563–574 DOI: 10.1541/ieejjia.22004793

Invited Paper

Machine Learning-based Remaining Useful Life Prediction Techniques

for Lithium-ion Battery Management Systems: A Comprehensive Review
Akash Samanta∗a) Non-member, Sheldon Williamson∗ Non-member

(Manuscript received May 12, 2022, revised Oct. 29, 2022)

J-STAGE Advance published date : March 3, 2023

Lithium-ion batteries (LIBs) are used to power a range of applications starting from portable consumer electronics to
electric vehicles and grid-tied energy storage systems. Now, with the increasing application of LIB in high power and
sophisticated applications, it is of great significance to predict the remaining useful life (RUL) for reliable operation
and to protect the battery pack from unwanted incidents including catastrophic failure. Real-time information on RUL
is essential to predict battery failure condition resulting in effective prevention or at least reduction of the damage that
may cause by the battery failure. Moreover, accurate RUL is extremely helpful for scheduling routine maintenance
and necessary replacement at the end of its useful life. Consequently, RUL prediction has become a topic of inter-
est to researchers. There are several RUL estimation techniques proposed in the last decade where machine learning
(ML)-based techniques showed superiority in terms of accuracy, adaptability, and modeling. Therefore, ML-based
RUL prediction methods are comprehensively reviewed based on their essential performance parameters in this paper.
A detailed discussion on the issues, challenges, trends, and future research scopes are also presented to provide clear
guideline to the researchers.

Keywords: prognostics and health management, electric vehicle, machine learning, deep learning, state estimation

Therefore, researchers have come up with different strate-

1. Introduction
gies for RUL prediction that can be broadly classified into
Energy storage technologies become predominant with the physical failure models and data-driven models. Recent ad-
increasing application of portable consumer electronics, elec- vancements and achievements in machine learning (ML), and
tric vehicles, and grid-tied energy storage systems for the in- data-driven ML-based techniques piqued interest in the state
tegration of renewable energy and load balancing. Lithium- estimation of LIB including RUL. An in-depth literature sur-
ion batteries (LIBs) are preferred over other battery technolo- vey indicates that researchers have proposed several ML-
gies owing to their longer cycle life, huge voltage density, based RUL prediction techniques in the last decade. The clas-
lighter weights, and less self-discharge. However, like other sification of RUL prediction techniques is shown in Fig. 1.
batteries, the capacity of LIB also degrades as the charge- Physical failure model-based methods can be further di-
discharge cycle increases. Therefore, the prediction of re- vided into two groups: empirical models (5)–(7) and
maining useful life (RUL) as part of the Prognostics and
Health Management (PHM) is extremely important to ensure
the reliability and safety of the battery system. Recent in-
cidents of catastrophic failure of the LIB system (1) (2) further
emphasized the utmost necessity of RUL prediction. Predic-
tion of RUL in advance not only ensures operational safety
while at the same time provides some key information about
the maintenance and replacement of batteries. The RUL of a
battery corresponds to the number of charges and discharge
cycles it can go through before reaching its end of life (EOL).
Where EOL generally corresponds to the time when the ca-
pacity of the battery is reduced below 70–80% of its nominal
capacity (3) . It is also noticed that LIBs used in EVs and in-
dustrial applications can reach their EOL much earlier than
intended because of fluctuating current loading, varying tem-
peratures, and stochastic depth of discharge (4) .
a) Correspondence to: Akash Samanta. E-mail: [Link]@
[Link]
∗
Department of Electrical, Computer and Software Engineering,
Ontario Tech University
Oshawa, ON L1G 0C5, Canada Fig. 1. Classification of RUL prediction techniques for LIB


c 2023 The Institute of Electrical Engineers of Japan. 563
 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

semi-empirical (8)–(10) . Semi-empirical methods map the degra-

dation parameters to some key battery parameters such as
static capacity, and impedance under different loading and
environmental condition through offline battery testing. Col-
lecting all the offline data requires lengthy and expensive
laboratory experiments which is one of the primary dis-
advantages of the semi-empirical model-based approaches.
Whereas empirical methods continuously update their model
parameters through online measurement/estimation of bat-
tery parameters and states. Typically, Bayesian filtering
methods such as Kalman filter (KF), particle filter (PF), and
their variants are used for online parameter updating simulta-
neously during battery cycling. Here, the prediction accu-
racy is sensitive to modeling accuracy. Now representing
the highly complex electrochemistry of LIB using a com- Fig. 2. General RUL Predict techniques for LIB
prehensive model is extremely challenging. Especially, ac-
curate representation of battery capacity fade phenomena us-
ing mathematical models such as the equivalent circuit model Neural Networks (ANN), Extreme learning machines (ELM),
(ECM) and electro-chemical model (EChM) is challenging as relevance vector machines (RVM), support vector machines
the parameters may vary with different environmental condi- (SVM), deep learning (DL), random forest regression (RFR)
tions. Therefore, ECM and EChM have very limited scope were used for RUL estimation of LIB. Thus, the existing RUL
in practical application due to poor robustness and low pre- prediction methods are grouped depending on the ML algo-
diction accuracy. Furthermore, EChM is limited by com- rithm used.
putational burden and difficult to implement in BMS with 2.1 Artificial Neural Network Techniques ANNs
limited processing power. PF, KF, sliding mode observer, are widely used due to their ability to discover nonlinear re-
and Lyapunov-based adaptive observer and their variants are lationships between variables, such as battery external pa-
also used with model-based methods for noise filtering to im- rameters to the battery degradation. The characteristic of
prove prediction accuracy. Model-based prediction methods LIB is highly influenced by temperature, discharge current,
demonstrated significant efficacy in describing the degrada- and the degradation process is also extremely nonlinear over
tion trend of batteries. However, complex calculations and the entire lifecycle. Thus, ANN was adopted to estimate the
a large number of parameters are often involved for accurate nonlinear degradation process due to its excellent capability
prediction and degradation modeling. That makes the model of nonlinear approximation. Previously, researchers started
inappropriate for practical application and real-time predic- building simple ANN for LIB state estimation however, as the
tion. Whereas the data-driven techniques first try to learn the RUL estimation is more complex and needs a more advanced
degradation pattern using historical data and predict future prediction model researchers started building hybrid models.
values based on current sets of input data. Data-driven tech- This is because it is noticed that the fusion of different models
niques can be broadly classified into statistical-based (11)–(16) provides better prediction results compared to a single model,
and machine learning (ML)-based (17)–(23) . ML-based methods especially in the case of complex nonlinear systems like LIB.
use external battery parameters to predict the battery state Cui and Hu (24) proposed a hybrid model of long-short terms
thus it is easier to implement and often minimizes the re- memory (LSTM) neural network and unscented Kalman fil-
quirement of accurate battery modeling and domain-specific ter (UKF) to capture nonlinear dynamics of LIB during both
experts. Typically, the change in capacity, voltage, coulom- charging and discharging periods while developing a state of
bic efficiency, impedance, and internal resistance are tracked health (SOH) and RUL prediction scheme. SOH is the ratio
by the ML algorithms to understand the degradation pattern of the current to the nominal capacity of the battery. LSTM
for RUL prediction. Furthermore, as the ML-based methods was first trained with the external battery parameters includ-
depend on the external battery parameters instead of model ing voltage, temperature, and current measurements to esti-
parameters thus the problem of poor dynamic accuracy of mate battery SOH then NN was employed to capture the ag-
model-based methods could be improved to a significant ex- ing mechanism of the LIB. Further UKF was also added to
tent. Realizing the potential future of ML-based strategies reduce the prediction error. The proposed method considered
for highly accurate and robust RUL prediction a detailed re- SOH as the primary indicator of RUL therefore the necessity
view of the state-of-the-art ML-based strategies is presented of two separate models for SOH and RUL is redundant. One
in this paper. The current issues, challenges, and future re- single LSTM model could have served the purpose. More-
search scopes are also highlighted for a ready reference and over, training and testing were conducted on the same data
baseline for further research and development. set. Therefore, the accuracy of prediction is questionable
with unknown sets of input data. The fusion of partial incre-
2. ML-based RUL Prediction Methods
mental capacity and ANN was employed by Zhang et al. (25)
The basic ML-based RUL prediction methods for LIB is for real-time analysis of battery data for RUL and SOH pre-
shown in Fig. 2. The state-of-the-art ML-based RUL pre- diction. This method used constant current discharge data for
diction techniques are discussed in the following section. ML-model training and testing which is far from the practical
Among the wide range of ML algorithms mostly Artificial scenario.

564 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

The aging data of the NASA Prognostics Center of Ex- Table 1. Summary of ANN-based RUL prediction tech-
cellence (PCoE) was massively used for SOH and RUL pre- niques
diction where four batteries are tested until the end-of-line
(EOL) criteria were reached. However, for developing a ro-
bust, accurate, and reliable model test data from only four
batteries is not sufficient. Therefore, an ANN-based RUL
prediction method was proposed by Jorge et al. (26) where a
new data set of battery aging, published by the department of
chemical engineering of the Massachusetts Institute of Tech-
nology was used. The data set consists of 124 cell aging data
from the beginning to EOL. The primary limitation of the
study is that battery aging, temperature variation, and fast
charging conditions were not considered. Moreover, collect-
ing such a large data set is suitable for experimental purposes. practical operation. Table 1 summarizes some of the major
It is a highly expensive and time-consuming process to col- ANN-based methods for RUL prediction.
lect such a large data set for all new battery sets. Shen et 2.2 Support Vector Machine (SVM)-based Strategies
al. (27) used a deep convolutional neural network (DCNN) in- SVM is another ML-based technique that aims to allevi-
stead of a basic ANN where voltage, current, and capacity of ate the necessity of obtaining model parameters and input
the discrete segment were used as input and deduced the un- information to the model in advance which is the primary
known capacity from the local charging cycle. Experiments limitation of model-based approaches for online prediction.
showed that this method has higher accuracy than SVM, but Often model-based approaches do not consider the uncertain-
the effect of temperature on capacity is ignored. Wu et al. (28) ties in practical operation and aging effects leading to inaccu-
proposed a feed-forward neural network (FFNN) and impor- rate prediction. However, the prognostics and health manage-
tance sampling (IS) method based online RUL prediction ap- ment of LIB demand fast monitoring and prediction of SOH
proach. ANN-model often falls into local optima because of and RUL. Earlier last decade, researchers like Zhou et al. (31) ,
over-training however the study did not consider any such Nuhic et al. (32) , and Patil et al. (33) started developing SVM for
special measures to handle the issue. RUL prediction. Thereafter, the researcher started upgrading
Liu et al. (29) developed an adaptive recurrent neural net- the earlier versions of simple SVM to more advanced SVMs
work (ARNN)-based RUL prediction scheme for LIBs where towards obtaining more accuracy and robustness in RUL pre-
the recursive Levenberg-Marquardt (RLM) method was used diction. However, almost all of the models did not consider
for adaptively optimize the network weights and biases for the aging effect and other uncertainties like temperature vari-
precise estimation over the entire life of the battery. De- ation in the prediction model. Thus, an online RUL predic-
spite the enhanced performance of ARNN compared to the tion method is realized by Zhou et al. (31) with the help of the
classical recurrent neural network (RNN) and the recurrent SVR algorithm and dynamic changes in battery parameters
neural fuzzy system (RNF) in RUL prediction, the adaptive are also incorporated. The selection of an appropriate kernel
architecture is very basic, and time-varying features are not function for SVR is the most important consideration, how-
considered completely during modeling. These may lead to ever, the selection of a mixed kernel for the study is not jus-
inaccuracy in the course of operation. Most of the RUL perdi- tified properly. The effectiveness was not demonstrated by
tion methods rely on the accurate estimation of the remain- conducting the comparative analysis of different kernel func-
ing useful capacity. However, the impact of capacity regen- tions with the same experimental data.
eration during the charging/discharging process significantly Commonly linear, polynomial, Gaussian, and sigmoid ker-
affects the accuracy of remaining useful capacity estimation. nels are used in SVM for mapping the data points from a
To alleviate this problem, Chen et al. (30) combined a Denois- lower to a higher dimensional space for accurate classifica-
ing Auto-Encoder (DAE) with the Transformer-based neural tion. Selection of the most appropriate kernel function for
network for RUL prediction where DAE was employed to a particular data set and the task is challenging. Therefore,
process the raw data. The model shows very accurate results Li et al. (34) conducted a comparison study among these com-
in relatively simple charge/discharge data. However, the per- monly used kernels and provided a guideline on the selection
formance of the model was not tested under dynamic operat- of a suitable kernel. The study showed that GA–SVR with a
ing conditions at varying operating temperatures which have radial basis function (RBF) and Particle Swarm Optimization
a large impact on degradation trends. Therefore, the predic- (POS)–SVR with an RBF produced more accurate prediction
tion accuracy in practical operations is questionable. Instead results. Abdelghafar et al. (35) demonstrated that Coyote Op-
of considering only the capacity fade trend for RUL predic- timization Algorithm (COA) with SVR performed better in
tion, Lui et al. (3) used a combination of measured voltage RUL prediction compared to conventional SVM. COA was
and capacity fade trend for RUL prediction. Experimental re- used for parameter optimization of SVM. The capacity-based
sults demonstrated that extrapolating the degradation param- approach consists of several limitations. The primary advan-
eters provided a more accurate and conservative RUL pre- tage of SVR is it can deal with the problems of small sample
diction when compared to simply extrapolating the capacity size, nonlinearity, and the analysis of time-series data. Wang
data only. Like other studies, the practical use scenarios were et al. (36) employed an artificial bee colony (ABC) for opti-
not considered in this study as well. Therefore, the accuracy mum parameter selection for SVR. Despite its improved per-
and robustness obtained in the laboratory do not guarantee in formance compared to POS-SVR, the study mostly focused

565 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

on the optimization algorithms compared to the improvement model, the internal resistance and cycle number of the battery
in SVR-based RUL prediction. were selected as health indicators. Despite good accuracy,
Du et al. (37) used SVR to establish a relationship between the model needs a large amount of training data and is prone
battery capacity, discharge power, and temperature with the to over-fitting resulting in poor efficiency and high computa-
RUL. Zhang et al. (38) enhanced the performance of their pre- tional time. Dong et al. (45) improved the model by implement-
viously developed RVM-based RUL prediction model by in- ing a multi-kernel support vector machine (MSVR) with the
tegrating a particle filter (PF) with RVM to construct the bat- PSO optimization technique. However, the information re-
tery degradation model. SVM and simple NN model suf- trieval efficacy of PSO is very poor and it easily gets stuck
fers from dimensionality problems and shallow architecture in local optima resulting ineffective selection of feature vec-
resulting in poor prognosis capability. ML-based methods tors and poor RUL prediction. To obtain the respective ad-
solely depend on historical degradation data. There is no re- vantages, RVM and PF, a fusion of data-driven methods and
quirement for an explicit mathematical model to represent the filter-based methods were used for RUL estimation by Saha
battery’s internal mechanism and degradation. Nuhic et al. (32) et al. (46) . RVM was used for internal parameter estimation
proposed an SVM-based SOH and RUL prediction approach. whereas PF was employed to find the optimum parameters of
Further, Patil et al. (33) improved the prediction accuracy by RVM. Zheng and Fang (47) also developed a similar approach
developing an online multi-stage SVM method where termi- but instead of PF, they employed an unscented Kalman fil-
nal voltage and temperature data were used as model input ter (UKF). Despite their superior performance in RUL pre-
parameters and characteristic parameters to improve the ac- diction, integration of multiple methods is difficult and often
curacy. However, these models may not guarantee high fore- increases the computational burden. Xu et al. (48) introduced a
casting accuracy due to the highly nonlinear characteristics normal particle swarm optimization method (NPSO) to bal-
of LIB. Chen et al. (39) SVR-based RUL prediction method ance the global search and local search capabilities. However,
with error compensation (EC). Firstly, capacity and discharge the improvement in prediction accuracy is minimal. More-
voltage differences are collected at equal time intervals to use over, the generalization capability of SVM-based methods is
as two health indicators. Then the ensemble empirical mode very poor.
decomposition (EEMD) is used for preprocessing the health Pearson correlation analysis was performed to extract the
indicator to minimize the influence of capacity regeneration feature vector (duration of an equal charging voltage differ-
and measurement noise. Phase space reconstruction (PSR) ence (DECVD)) from charging voltage data by Gou et al. (49) .
with the C–C technique was used to design the EC mecha- Then, an ensemble learning method was used for RUL pre-
nism. PSO was also used for the optimum parameterization diction. The Random Vector Functional Link (RVFL) net-
of the SVR model. The primary reason for improved accu- work was used to develop a mapping relation between the
racy is the integration of EC. However, health indicators are feature vector and battery capacity. The RVFL has a much
not adaptive to battery aging and capacity regeneration. faster learning rate, better generalization capability, and a
Chen et al. (40) developed an integration algorithm that uses computational cost-tuning mechanism. The learning speed of
battery external parameters and SVM for RUL prediction that RVFL can be thousands of times faster than traditional learn-
helps to achieve a better decision. Liu et al. (41) proposed an ing algorithms (such as SVM) and requires much less compu-
RVM-based RUL prediction algorithm with a flexible and ef- tation. The performance and useful capacity of LIB degrade
fective online training strategy. However, RVM is suitable over time due to aging and operating conditions. Therefore,
for small-scale training data. Xu et al. (42) combined a cata- determining the SOH and RUL of LIB is essential for safe
clysmic mutation genetic algorithm (CMGA) and SVM for and reliable operation. The SOH represents the remaining
RUL prediction of LIB which demonstrated improved per- effective capacity whereas RUL provides tentative informa-
formance compared to a genetic algorithm (GA)-SVR and tion about the end of the battery’s useful life. Accurate in-
Cross-Validation (CV)-SVR models. Despite the fundamen- formation on SOH is the prerequisite of the RUL prediction.
tal building block of RUL prediction was still SVR, the im- Thus, Gou et al. (50) proposed a hybrid random learning tech-
proved performance was primarily due to the use of CMGA nique by integrating an extreme learning machine (ELM) al-
for parameter optimization of the SVR model. The com- gorithm and a random vector functional link (RVFL) network
parative analysis showed that the proposed method reduced for SOH and RUL prediction. The training speed of ELM
the average Mean squared error (MSE) by 0.029 and 1.608 and RVFL is much higher than the traditional learning algo-
whereas the average squared correlation coefficient (R2 ) in- rithms. The randomness enabled ensemble learning that im-
creased by 0.8% and 4.4%, respectively. Considering the proved the generalization capability of the proposed method.
impact of capacity regeneration on the RUL prediction is a The nonlinear autoregressive (NAR) structure helped to cap-
challenging issue and ignoring the effect leads to significant ture the historical dependence resulting in improved predic-
inaccuracy in RUL prediction over the long time frame. Here, tion accuracy. At first, SOH is predicted using the present
Xu et al. (42) stated that they have considered the influence of external measurements and past SOH information. Further,
capacity regeneration, measurement noise, and dynamic op- NAR is used to predict the RUL. The determination of model
erating conditions on RUL prediction. However, it was not robustness was not demonstrated properly. Similarly, ELM
clearly explained and demonstrated. Wang et al. (43) and Wei and PF were used by Sun et al. (51) where PF was used for
et al. (44) developed SVR-based RUL prediction methods with noise filtering.
different optimization techniques to find the SVR parame- RUL prediction based on health status monitoring was
ters. The former study used artificial bee colony (ABC) opti- proposed by Xue et al. (52) where Random forest regres-
mization whereas PSO was used in the latter study. In Wei’s sion (RFR) was exploited for capacity estimation and the

566 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

Box-Cox transformation (BCT) method was used for a linear Table 2. Summary of SVM-based RUL prediction tech-
mapping of capacity and RUL. Besides charging/discharging niques
data, the consideration of thermal characteristics in battery
capacity estimation enhanced the accuracy of RUL predic-
tion. Although, linear mapping between battery capacity and
RUL is considered the transformed capacities are not strictly
linear with the cycle number, which eventually leads to inac-
curacy in RUL prediction. The reported Mean Absolute Er-
ror (MAE) in RUL prediction remains within 40 cycles with
a 95% confidence boundary.
As capacity is an important indicator of RUL, thus Abdel-
ghafar et al. (35) proposed a combined Coyote Optimization
Algorithm (COA) with a Support Vector Regression (SVR)-
based capacity estimation approach for RUL prediction. Fur-
ther, a comparative analysis also showed that the COA-SVR-
based approach is more effective compared to the conven-
tional SVR algorithm with randomized parameter selection
and a relevance vector machine (RVM). Where the COA
looks for optimal parameters in the prediction region for the
SVR model making the model adaptive to aging parameters.
As an improvement to the single SVM models, Xue et al. (53)
proposed a hybrid model of adaptive unscented Kalman fil-
ter (AUKF), Genetic algorithm (GA), and Support vector re-
gression (SVR) that demonstrated a much better filtering ef-
fect compared to single unscented Kalman filter (UKF) and
a combination of RVM-UKF (47) methods. Recently, Wang et
al. (54) conducted a comparative study among quantum parti-
cle swarm optimization – support vector regression (QPSO–
SVR) and particle swarm optimization and support vector re-
gression (PSO–SVR), and single support vector regression
(SVR) to demonstrate that the fusion of different methods
performs better compared to single algorithms. Table 2 sum-
marizes some of the major SVM-based methods for RUL pre-
diction.
2.3 Relevance Vector Machine (RVM) RVM was
introduced by Tipping in 2001 (55) and is gaining importance Table 3. Summary of RVM-based RUL prediction tech-
niques
in prognostics. Furthermore, based on Bayesian theory, RVM
offers probabilistic outputs. It has sparser solutions and faster
performance on test data than SVM whilst maintaining com-
parable generalization errors. Zhao et al. (56) combined RVM
and the gray model (GM) for RUL prediction. However, the
model completely relies on the historical data that did not ac-
count for the future changes in battery parameters that are
subject to change with the operating conditions. Guo et al. (57) prediction. The rest of the algorithms are optimized and
first extracted fourteen battery health indicators using princi- evolved from DBN and RNN. However, later it was realized
pal component analysis from charging data for capacity esti- that simple RNN is not suitable to obtain long-term depen-
mation. Finally, RVM was used for RUL prediction. This dence from time-series data. Moreover, RNN suffers from
method encompasses similar drawbacks to capacity-based a vanishing gradient problem. Thus, researchers developed
RUL prediction methods. Table 3 summarizes some of the more advanced DL architecture such as LSTM. LSTM is ca-
major RVM-based methods for RUL prediction. pable to acquire long-term dependencies in time-series pre-
2.4 Deep Learning-based Approach In comparison diction such as SOH and RUL of LIB. Bayesian model aver-
to the conventional neural network, deep neural networks aging (BMA), an ensemble method was used by Liu et al. (58)
(DNN) are capable to learn high-dimensional system char- to capture uncertainties and further used LSTM for RUL pre-
acteristics from the input data. Only a simple preprocessing diction. Typically, voltage, current, and temperature during
is typically required before input to the model. DL-based overcharging and discharging cycles with different working
approaches eventually eliminate the requirement of complex conditions are used for extracting key features for health es-
and time-consuming feature extraction and dimensionality timation. These data are then used to train and build the RUL
reduction of data. Deep Belief Networks (DBN) and Recur- prediction model. Recently, A hybrid model of RNN and
rent Neural Networks (RNN) are the primary variants of DL LSTM (RNN-LSTM) was used by Chinomona et al. (59) for
algorithms suitable for time series prediction such as RUL RUL prediction and demonstrated highly accurate prediction

567 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

with root mean square error (RMSE) of 0.00286 and mean Despite hybrid approaches demonstrating improved per-
average error (MAE) of 0.00222. The algorithm effectively formance in real-time prediction, determining the model pa-
selects the significant feature subset resulting in improved ac- rameters is often too complex. Huang et al. (68) used the Box-
curacy with a lower amount of the training data requirement Cox transformation (BCT) for model parameter optimiza-
and improved processing speed. Despite high prediction ac- tion and transforming the nonlinear function into a linear
curacy in partial battery discharge data, it is inclusive that the one. Then a mapping relation between energy-throughput
proposed method could provide satisfactory results over the and RUL was established using ANN. The initial input data
entire life cycle of the battery. was label-free thus, this method reduced the requirement of
Qiao et al. (60) used LSTM to capture the capacity regener- labeled data reducing the cost of data collection and time of
ation phenomena of the battery. Further, DNN was used to data collection. BCT transforms the irregular data into a lin-
predict the global deterioration trend in the RUL prediction. ear shape for highly accurate prediction. However, the trans-
Song et al. (61) used a hybrid model of CNN and LSTM to es- formed parameter differs significantly from the variation in
tablish a relationship between battery SOC, SOH, and RUL. the data length. For that reason, complete life-cycle data is
Li et al. (62) used a deep belief network to learn battery capac- required but collecting this data beforehand without reach-
ity degradation to estimate SOH. Later the estimated SOH ing RUL is not possible. To avoid these issues, time-series
was used for predicting RUL. A temporal convolutional net- prediction methods are evolved that use sparsity selection to
work (TCN)-based SOH monitoring model framework was find dominant data points for developing degradation mod-
proposed by Zhou et al. (63) where Causal convolution and di- els. The sparsity selection also helps to reduce the prediction
lated convolution techniques were employed to capture local error.
capacity regeneration which is one of the major problems. Li et al. (69) developed a fusion method where a combination
In addition, empirical mode decomposition (EMD) technol- of the Elman neural network (ENN) and LSTM for RUL pre-
ogy was used to denoise the offline training data. Further, to diction under high and low-frequency sub-layers. A dilated
avoid overfitting, another primary issue of DL-based meth- CNN-based neural network was used for RUL prediction by
ods residual connection, and dropout technologies were used. Hong et al. (70) where model features were automatically cho-
These improved the training speed but they have not consid- sen by the algorithm from the input data. The study demon-
ered the long-term regeneration. Li et al. (64) upgraded the strated that the model can not only capable to capture the
conventional structure of LSTM to AST-LSTM and the ef- long-term temporal patterns in the battery data while at the
fectiveness was demonstrated for RUL and SOH prediction. same the prediction is (25X) faster and more accurate (MAE
Unlike conventional LSTM structures, the input and forget = 10%) compared to state-of-the-art RUL prediction methods
gates are coupled by a fixed connection in the AST-LSTM. in its class. Very limited cycle data is sufficient for the pro-
The modified architecture helps to simultaneously determine posed model to learn the aging characteristics of LIB. The
old and new data. Furthermore, the element-wise product of consideration of inter-cycle aging in the prediction model
the new and historical data alongside a peephole connection was the primary reason for the good accuracy. The predic-
from the constant error carousel (CEC) improved the predic- tion error is still not suitable for practical applications thus the
tion results and filter out the unwanted error signals. How- prediction accuracy needs to be improved. Moreover, prepro-
ever, the training and development of the model were highly cessing of data was not done, which could lead to significant
complicated and computationally inefficient. uncertainty in a real-world application. Khumprom et al. (71)
An adaptive RNN based on the recursive Levenberg- developed a very basic DL-based approach. The testing and
Marquardt method was developed by Liu et al. (29) to under- validation data and process were incomplete and very basic.
stand the trend of impedance degradation for mapping be- Traditional LSTM models realize a one-to-one mapping
tween impedance degradation and RUL. Qu et al. (65) used between the input and output layer which is not that efficient
LSTM and PSO for RUL prediction where PSO was used for and flexible. Therefore, Park et al. (72) developed an LSTM
model parameter optimization. Further, Saha et al. (46) pro- model with many-to-one mapping for RUL prediction. Exter-
posed an autoregressive and moving average (ARMA) for nal battery measurements including voltage, current and tem-
time series data analysis such as RUL prediction. The slid- perature, and charging profiles and their variation with aging
ing window sizes (SWS) in DL models are often defined em- were considered as model input parameters. The many-to-
pirically which is time-consuming and may not be the opti- one structure increases flexibility and substantially reduces
mum always. Therefore, Ma et al. (66) applied the false near- the number of model parameters for better generalization.
est neighbor method to determine SWS and demonstrated its Short-term regeneration was considered but the effect of cell
impact using one DL-based RUL prediction method. Ma aging was not considered. Traditional ML-based strategies
et al. (66) utilized the false nearest neighbors method for de- are not adaptive to the changes in operating and environmen-
termining the optimal sliding window size of a hybrid RUL tal conditions leading to cascading errors with the increase in
prediction method that combines CNN and LSTM. The ap- battery cycle aging. To remedy this, Che et al. (73) proposed
proach also offered wide generality and enhanced prediction a self-correcting RUL prediction scheme combining trans-
accuracy. The combined approach can alleviate the issues of fer learning and a gated recurrent neural network (GRNN).
individual CNN and LSTM-based methods. An investigation Determination of end-of-life (EOD) in the online prediction
on DL and auto-encoder with deep NN-based RUL prediction of RUL is one of the primary challenges. Thus, Gaussian
was conducted by Ren et al. (67) . The method demonstrated process regression is used to determine the optimum value
high accuracy, but the model development method was not of EOL, and the online model correction feature is imple-
clearly explained. mented using the transfer learning method. The model was

568 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

fine-tuned with early degradation cycling data to provide ac- Table 4. Summary of deep learning-based RUL predic-
curate predictions throughout the life cycle. Inaccurate pre- tion techniques
diction during early states is a serious concern in the case of
ML-based prediction strategies, here the self-correcting net-
work demonstrated very good performance throughout the
life cycle in RUL prediction. This study first considered the
extraction of health indicators under fast charging conditions
however the major concern is that it considered SOH as the
ratio of current capacity and nominal capacity which is not a
very accurate way to predict RUL.
Primarily most of the prediction methods forecast the RUL
by checking the remaining useful capacity of the battery.
Zraibi et al. (21) proposed a hybrid model of LSTM, and DNN
termed for RUL prediction. Besides improvement in the pre-
diction accuracy compared to simple CNN-LSTM, the im-
provement in the computational time (data extraction, data
formatting, training, validation, and performance computing)
was also noticed. However, the model still used SOH as the
feature vector for calculating the RUL of the battery, there-
fore, encompassing similar drawbacks. Moreover, the effect
of variation in operating temperature was also not consid-
ered. Zhang et al. (74) used a hybrid model of LSTM-RNN for
RUL prediction. A combination of the deep belief network
(DBN) and LSTM network was used by Cao et al. (75) for RUL
prediction and showed that the combined approach is better
compared to DBN and LSTM alone for the same task. Here,
still, the capacity data was used as an input feature. There- Table 5. Summary of extreme learning machine-based
fore, a similar problem persists. The proposed DBN-LSTM RUL prediction techniques
model was very simple with only a single hidden layer which
was not sufficient to capture the complex aging behavior of
LIB. They have considered the local regeneration phenom-
ena for the rest state of the battery, the regeneration phenom-
ena under dynamic operating conditions were not considered
which is more practical. Moreover, handling the overfitting
and underfitting problems were not been addressed by any of
Table 6. Summary of RFR-based RUL prediction tech-
the above-mentioned studies. These problems are primarily
niques
due to the improper selection of model parameters such as
the number of hidden layers and nodes. Researchers must
need to put some effort into it in future studies. Table 4 sum-
marizes some of the major deep-learning-based methods for
RUL prediction.
2.5 Extreme Learning Machine-based Approach prediction by Chen et al. (78) . The SGM was used to model the
Aiming to minimize the issues of trapping in local op- capacity degradation of the battery. The hybrid model can ex-
tima, simultaneous operation of prediction, and parameter trapolate the rate of change of capacity to realize RUL using
optimization models, Fan et al. (76) proposed a Forgetting On- very less data points. The aging behavior was studied under
line Sequential Extreme Learning Machine (FOS-ELM) for constant discharging current. The effect of dynamic loading
RUL prediction. Further, the Hybrid Grey Wolf Optimizer with the varying working environment was not considered.
(HGWO) algorithm was employed for model parameter opti- Thus, the reported reliability and accuracy are questionable
mization and Variational Mode Decomposition (VMD) for in a practical scenario. Moreover, the effect of depth-of-
input noise filtering. The prediction performance and ef- discharge and energy recovery was not considered. Table 5
ficiency were better compared to standard RNN, standard summarizes some of the major extreme learning machine-
LSTM, and RVM. Zhu et al. (77) combined extreme learn- based methods for RUL prediction.
ing machine (ELM), gray wolf optimization (GWO), and 2.6 Random Forest Regression (RFR) Li et al. (79)
differential evolution (DE) and significantly improved the proposed a random forest regression (RFR)-based RUL pre-
RUL prediction accuracy. One of the primary limitations diction method where partial charging voltage data was used
of ML-based approaches is the limitation of model training to extract feature vectors to estimate the battery capacity pre-
and initialization data. Insufficient data leads to inaccuracy cisely. Partial charge/discharge data is not sufficient to model
and unreliability of the prediction model. To partially mini- the battery capacity and degradation mechanism. Thus, accu-
mize the problem, a combination of the sliding-window grey racy cannot be guaranteed. Table 6 summarizes some of the
model (SGM) and particle filter (PF) was proposed for RUL major RFR-based methods for RUL prediction.

569 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

lutions here. The issue of inaccurate measurement of SOH

3. RUL Prediction of Second-Life LIB
using health indicators due to partial discharge of cells in a
As discussed earlier, the EOL is defined as the reduction discharge cycle results in an error in RUL prediction. There-
of 20%–30% of the nominal capacity. From here it could be fore, instead of estimating SOH first using health indicators,
inferred that the retired LIBs may not be suitable for EV ap- directly estimating RUL using health indicators may increase
plications however the batteries remain a significant amount the prediction accuracy. Furthermore, effective extraction of
of usable capacity for several low-demanding applications in health indicators under fast charging conditions is also an-
terms of power requirement and depth-of-discharge such as other challenge. These also need further research and con-
on-grid applications. The reuse of these retired batteries in- sideration.
stead of recycling is termed as second-life LIB. The second SVR demonstrated excellent performance in estimation
life of LIBs has been considered highly economical, and en- and prognosis however it is best for small sample data set.
vironmentally friendly (80) . A simple ML-based classification SVR can obtain the best-fitting linear curve by minimizing
algorithm was used by Sanz-Gorrachategui et al. (81) for clas- the structural risk however the clustering is not optimum with
sifying the RUL of a second-life LIB as “long RUL” or “short the increasing number of data set and feature vector. Despite
RUL” classes. Further, they have also developed a multi- the acceptable performance of SVM, RVM, Gray model, and
variable linear regression (MLR)-based exact RUL prediction other hybrid approaches, their accuracy cannot be guaranteed
method with considerable accuracy. The primary concern of for all life cycle stages of the battery. Often, basic ML-based
the study is that it used data from several pristine LIB and schemes fail to extract the valuable information hidden in-
preprocessed the data to use for the second-life battery RUL side the historical data, and the accuracy of experimentally
prediction. This data may not represent the actual behavior of training and validated models cannot be guaranteed due to
the retired battery. Proper experimental tests of the second- the extreme nonlinear characteristics of LIB. Irrespective of
life battery should be conducted for modeling. estimation techniques, the exact value of maximum battery
capacity is required for RUL prediction. Now, it is difficult
4. Issues, Challenges, and Future Research to quantify the actual maximum capacity without discharging
Scopes in ML-based RUL Prediction of LIB the battery completely. Therefore, an online determination is
Compared to model-based RUL prediction methods, ML- highly challenging and inaccurate. Therefore, it is recom-
based methods require a large set of offline training data that mended to obtain a mapping relation between the partial dis-
increases the data collection time, experimental cost, as well charge capacity to the full discharge capacity beforehand for
as computational cost. To remedy this, few studies used par- practical purposes. However, it should be ensured that the im-
tial battery degradation data for model training. However, the pact of battery aging and temperature on the battery capacity
degradation rate is much more rapid in the later phase com- is well taken care of.
pared to the early phase of the battery life resulting in sig- In practice, discharge currents vary randomly that have a
nificant inaccuracy in RUL prediction using the ML model, greater impact on battery life. Most of the state-of-the-art
trained using partial lifecycle data. prediction methods rarely considered this option seriously.
EOL is typically defined as the loss of 20%–30% of its Predicting RUL under variable discharge current needs to
nominal capacity and RUL is the difference between EOL be explored. The global degradation process refers to ideal
and the current cycle number. Now estimating the battery aging and does not consider the capacity regeneration phe-
capacity in real-time is difficult due to incomplete discharge nomenon. Whereas local degradation is considered the re-
in each cycle. Therefore, typically different health indica- generation phenomenon which is inevitable in all practical
tors show regular variations with battery aging. Health indi- LIB. The consideration of global and local aging process
cators generally are extracted from the charging/discharging needs to be considered for developing the practical purpose
data to estimate the SOH. Now the selection of health in- RUL prediction method.
dicators for SOH estimation and the selection of ML algo- ML models require large historical data sets for training
rithm for regression model training usually affect the accu- and the prediction input vectors should follow the same pat-
racy of RUL prediction. Optimum selection of health in- tern for an accurate prediction. Such an intelligent model
dicators till remains challenging. In addition, the capacity needs to be developed for accurate prediction under the prac-
regeneration phenomena significantly affect the RUL predic- tical scenario. Moreover, training data sets are typically ob-
tion accuracy. Another primary shortcoming of conventional tained from the initial degradation stage which is much differ-
ML algorithms including SVM, RVM, and GPR is the van- ent compared to the later stage of the battery. Thus, this kind
ishing gradient problem. Researchers therefore now started of model does not guarantee an accurate prediction through-
using DL-based methods including RNNs, long short-term out the life cycle of the battery.
memory (LSTM), and gated recurrent unit (GRU) to capture Typically, a LIB undergoes three stages during its actual
the long-term dependencies. Still, determining RUL based on useful life cycle namely, charging, discharging, and rest. Bat-
SOH leads to error accumulation due to inaccuracy in SOH tery aging under rest conditions was not deeply studied. Dur-
prediction. Furthermore, batteries often experience different ing the rest period, some energy is restored automatically
degradation patterns under different operating conditions and which is commonly known as regeneration which signifi-
even the aging of different battery cells in a battery pack is cantly affects the accuracy of RUL prediction. Moreover,
different. It remains still challenging how to optimally uti- regeneration phenomena are of two types, charge regenera-
lize historical information. Developing a robust and intel- tion which may result after a very short rest period (a few
ligent adaptive algorithm could be one of the potential so- minutes), and capacity regeneration which results after a few

570 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

hours or even after a few days. Therefore, distinguishing and nature of the LIB chemistry and aging phenomena, the opti-
considering these phenomena is challenging. These need fur- mization algorithms need to be run simultaneously. Operat-
ther research and understanding. Predicting RUL under the ing both models in parallel may lead to modeling complex-
presence of regeneration phenomena needs to be deeply stud- ity and computational burden but that needs to be considered
ied. separately.
A large amount of training data is extremely essential for RNN and LSTM suffer from underfitting and low predic-
developing a high-fidelity ML model. Now collecting such a tion accuracy due to limited battery data and noisy data.
large data set is time-consuming and expensive. Thus, often Moreover, in practical application, RNN-based prediction
partial degradation data is used however the degradation rate methods become unreliable due to the presence of capacity
significantly varies throughout the entire battery life. Such a regeneration. These methods can only provide accurate pre-
model could not accurately track the aging trend of the bat- diction if a relaxation period is permitted to the battery after
tery leading to poor RUL prediction accuracy and robustness. use. Therefore, online prediction is erroneous. CNN is de-
The prediction accuracy of ML-based methods strongly de- signed to alleviate the problem of under-fitting phenomena
pends on the accuracy of training data. However, collecting due to limited and noisy data. The characteristics data used
all battery aging data considering all uncertainties in the lab- during ML-model training changes by the time of battery use,
oratory environment is highly challenging. SVM performed thus training data obtained in the laboratory do not accurately
better compared to ANN and ARMA, especially during the represent the battery characteristics during practical applica-
later cycle life of a battery. RVM has the potential to cap- tion. Thus, a direct prediction method with real-time data
ture uncertainties however the results are not stable due to the needs to be developed.
dynamic volatility of capacity measurements. In DL-based Despite, second life (retired) LIBs can be used in a range
methods, the proper learning process requires a large amount of applications, the RUL prediction of retired batteries using
of training data, however, sometimes overfitting problems ML-based methods has not yet been seriously studied. The
arise. Therefore, to alleviate the issues like overfitting, re- additional challenge of ML-based RUL prediction of retired
searchers started using techniques like dropout (randomly batteries is the absence of historical usage information (his-
drop units and connection) and L1&L2 regularization. The torical data). Researchers should consider this in mind as
optimum selection of these ML parameters is hard to deter- well. A summary of the advantages and limitations of differ-
mine. These need further research and development. ent ML-based strategies is presented in Table 7.
It is worth noting that most of the research studies fo-
5. Conclusions
cused on improving the prediction accuracy and efficiency
of the RUL prediction however, very few focused on the op- This paper discussed in detail the importance of RUL pre-
timum determination of health features. Thus, the extraction diction for the reliable and safe operation of LIB which is
of health feature vectors needs to be explored in more de- commonly used in a large range of applications starting from
tail as these are the backbones of any ML-based strategies. consumer electronics to electric vehicles and stationary grid-
The local regeneration phenomenon that is self-charging of tied energy storage systems. State-of-the-art RUL predic-
LIB during the rest period greatly impacts the starting point. tion strategies are logically classified, and the superiority of
Now, the capacity-based RUL prediction scheme typically ML-based strategies over other strategies is explained clearly.
determines the next capacity based on its previous predic- That further justified the importance of emphasizing the ML-
tion. Now, if the initial prediction point has some error, then based RUL prediction strategies in this paper. The paper
the error increases cumulatively over the cycle. To avoid such thoroughly reviewed the existing ML-based RUL prediction
issues, direct prediction techniques need to be developed. methods and systematically presented a clear picture of the
During the charging and discharging process, several baseline. It is noticed from the critical analysis of the existing
noises are generated in the measurement signal due to the RUL prediction methods that at the initial phase, researchers
complex physical and electrochemical changes in the battery. started developing SVM and SVR-based techniques. Slowly
These noises greatly influence prediction accuracy. Exist- ANN and the hybrid models of NN are developed. Recently,
ing denoising methods such as KF, PF, empirical mode de- the DL and other advanced techniques are introduced in the
composition (EMD), ensemble empirical mode decomposi- RUL prediction domain. It is noticed from the survey that
tion (EEMD), and complete ensemble empirical mode de- the primary shortcoming of the data-driven approaches is the
composition with adaptive noise (CEEMDAN) were used for necessity of large historical data sets which are typically not
model-based techniques mainly. These techniques or similar available for research and development purposes. Another
approaches must need to be integrated with ML-based strate- major issue is the poor robustness leading to inaccurate long-
gies for noise filtering and improving prediction accuracy. term prediction. Besides computational cost, developing a
In most of the ML-based techniques, although the algo- high-fidelity model that can effectively learn the degradation
rithms contain many feature vectors, however, often the RUL trend of the battery is challenging. Often the capacity regen-
prediction primarily depends on a single key parameter such eration phenomenon was ignored that must need to be consid-
as capacity degradation. Moreover, finding the key param- ered carefully. Although DL and advanced hybrid strategies
eters from battery test data is also challenging. Thus, dif- are currently new, these have the potential to be used in real-
ferent optimization algorithms such as differential evolution life applications. DL models, especially hybrid models are
(DE), grey wolf optimizer (GWO), particle swarm optimiza- associated with a high computational cost that also needs fur-
tion (PSO), and genetic algorithm (GA) were used for finding ther modifications to improve the computational speed and
and adjusting key parameters. However, due to the dynamic efficiency. Instead of doing research primarily on efficacy

571 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

Table 7. Summary of the advantages and limitation of

ML-based strategies References

(1) Q. Wang, P. Ping, X. Zhao, G. Chu, J. Sun, and C. Chen: Thermal runaway
caused fire and explosion of lithium ion battery (2012)
(2) H. Tian, P. Qin, K. Li, and Z. Zhao: “A review of the state of health for
lithium-ion batteries”, Research status and suggestions (2020)
(3) Y.H. Lui, M. Li, A. Downey, S. Shen, V.P. Nemani, H. Ye, C. VanElzen,
G. Jain, S. Hu, S. Laflamme, and C. Hu: “Physics-based prognostics
of implantable-grade lithium-ion battery for remaining useful life predic-
tion”, J. Power Sources., Vol.485, p.229327 (2021) [Link]
[Link].2020.229327
(4) X. Hu, S.E. Li, and Y. Yang: “Advanced Machine Learning Approach for
Lithium-Ion Battery State Estimation in Electric Vehicles”, IEEE Trans.
Transp. Electrif (2016) [Link]
(5) X. Su, S. Wang, M. Pecht, L. Zhao, and Z. Ye: “Interacting multiple model
particle filter for prognostics of lithium-ion batteries”, Microelectron. Reliab
(2017) [Link]
(6) L. Zhang, Z. Mu, and C. Sun: “Remaining Useful Life Prediction for
Lithium-Ion Batteries Based on Exponential Model and Particle Filter”, IEEE
Access., Vol.6, pp.17729–17740 (2018) [Link]
2018.2816684
(7) D. Shen, L. Wu, G. Kang, Y. Guan, and Z. Peng: “A novel online method
for predicting the remaining useful life of lithium-ion batteries consider-
ing random variable discharge current”, Energy, Vol.218, p.119490 (2021)
[Link]
(8) S. Sankararaman, M.J. Daigle, and K. Goebel: “Uncertainty quantification in
remaining useful life prediction using first-order reliability methods”, IEEE
Trans. Reliab (2014) [Link]
(9) C. Lyu, Q. Lai, T. Ge, H. Yu, L. Wang, and N. Ma: “A lead-acid bat-
tery’s remaining useful life prediction by using electrochemical model in
the Particle Filtering framework”, Energy (2017) [Link]
[Link].2016.12.004
( 10 ) F. Yang, X. Song, G. Dong, and K.L. Tsui: “A coulombic efficiency-based
model for prognostics and health estimation of lithium-ion batteries”, Energy,
Vol.171, pp.1173–1182 (2019) [Link]
( 11 ) L. Chen, J. An, H. Wang, M. Zhang, and H. Pan: “Remaining use-
ful life prediction for lithium-ion battery by combining an improved
particle filter with sliding-window gray model”, Energy Reports (2020)
[Link]
( 12 ) A. Mosallam, K. Medjaher, and N. Zerhouni: “Data-driven prognostic
method based on Bayesian approaches for direct remaining useful life pre-
diction”, J. Intell. Manuf (2016) [Link]
( 13 ) H. Feng and G. Shi: “SOH and RUL prediction of Li-ion batteries based on
improved Gaussian process regression”, J. Power Electron (2021) [Link]
org/10.1007/s43236-021-00318-5
( 14 ) D. Liu, Y. Luo, J. Liu, Y. Peng, L. Guo, and M. Pecht: “Lithium-ion bat-
tery remaining useful life estimation based on fusion nonlinear degradation
AR model and RPF algorithm”, Neural Comput. Appl., Vol.25, pp.557–572
(2014) [Link]
( 15 ) S.S.Y. Ng, Y. Xing, and K.L. Tsui: “A naive bayes model for robust remaining
useful life prediction of lithium-ion battery”, Appl. Energy, Vol.118, pp.114–
123 (2014) [Link]
( 16 ) K. Pugalenthi, H. Park, S. Hussain, N. Raghavan: “Remaining Useful Life
improvement, the selection of health indicators (feature vec-
Prediction of Lithium-Ion Batteries Using Neural Networks with Adaptive
tors) needs to be studied. The accuracy of SVM and GPR- Bayesian Learning”, Sensors, Vol.22, p.3803 (2022) [Link]
based strategies is significantly depending on the optimum s22103803
selection of kernel function and parameters. A clear guide- ( 17 ) T. Tang and H. Yuan: “A hybrid approach based on decomposition algo-
rithm and neural network for remaining useful life prediction of lithium-ion
line behind the selection needs to be developed. It is also battery”, Reliab. Eng. Syst. Saf. (2022) [Link]
noticed from the study that the limited availability is one of 108082
the primary reasons behind the limited generalization abil- ( 18 ) R. Liu: “Remaining Useful Life Prediction of Lithium-ion Batteries Us-
ing Multiple Kernel Extreme Learning Machine”, Recent Adv. Comput. Sci.
ity of the existing methods. A large set of battery data from
Commun. (2020) [Link]
all major cell manufacturers and especially the historical use ( 19 ) Z. Chen, N. Shi, Y. Ji, M. Niu, and Y. Wang: “Lithium-ion batteries remain-
data is important to design and validate the advanced tech- ing useful life prediction based on BLS-RVM”, Energy, Vol.234, p.121269
niques. There is a massive potential for the potential use of (2021) [Link]
( 20 ) M.A. Patil, P. Tagade, K.S. Hariharan, S.M. Kolake, T. Song, T. Yeo, and S.
second-life LIBs. However, the RUL prediction strategies for Doo: “A novel multistage Support Vector Machine based approach for Li ion
second-life LIBs is not yet been explored considerably. This battery remaining useful life estimation”, Appl. Energy, Vol.159, pp.285–297
needs to be explored urgently. Finally, the effect of cell aging (2015) [Link]
and the various operating conditions including the tempera- ( 21 ) B. Zraibi, C. Okar, H. Chaoui, and M. Mansouri: “Remaining Useful
Life Assessment for Lithium-Ion Batteries Using CNN-LSTM-DNN Hybrid
ture variations are not considered in the RUL prediction so Method”, IEEE Trans. Veh. Technol., Vol.70, pp.4252–4261 (2021) https://
far. That must be considered in future developments for bet- [Link]/10.1109/TVT.2021.3071622
ter realistic predictions. ( 22 ) K.S.R. Mawonou, A. Eddahech, D. Dumur, D. Beauvois, and E. Godoy:

572 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

“State-of-health estimators coupled to a random forest approach for lithium- ( 42 ) J. Xu, Y. Ni, and C. Zhu: “Remaining Useful Life Prediction for Lithium-
ion battery aging factor ranking”, J. Power Sources (2021) [Link] Ion Batteries Based on Improved Support Vector Regression. Diangong Jishu
10.1016/[Link].2020.229154 Xuebao/Transactions China Electrotech”, Soc., Vol.36, pp.3693–3704 (2021)
( 23 ) J. Liu and Z. Chen: “Remaining useful life prediction of lithium-ion batteries [Link]
based on health indicator and Gaussian process regression model”, IEEE Ac- ( 43 ) Y. Wang, Y. Ni, S. Lu, J. Wang, and X. Zhang: “Remaining Useful Life Pre-
cess, Vol.7, pp.39474–39484 (2019) [Link] diction of Lithium-Ion Batteries Using Support Vector Regression Optimized
2905740 by Artificial Bee Colony”, IEEE Trans. Veh. Technol., Vol.68, pp.9543–9553
( 24 ) X. Cui and T. Hu: “State of Health Diagnosis and Remaining Useful Life (2019) [Link]
Prediction for Lithium-ion Battery Based on Data Model Fusion Method”, ( 44 ) J. Wei, G. Dong, and Z. Chen: “Remaining Useful Life Prediction and State
IEEE Access, Vol.8, pp.207298–207307 (2020) [Link] of Health Diagnosis for Lithium-Ion Batteries Using Particle Filter and Sup-
ACCESS.2020.3038182 port Vector Regression”, IEEE Trans. Ind. Electron., Vol.65, pp.5634–5643
( 25 ) S. Zhang, B. Zhai, X. Guo, K. Wang, N. Peng, and X. Zhang: “Synchronous (2018) [Link]
estimation of state of health and remaining useful lifetime for lithium-ion bat- ( 45 ) D. Gao and M. Huang: “Prediction of remaining useful life of lithium-ion
tery using the incremental capacity and artificial neural networks”, J. Energy battery based on multi-kernel support vector machine with particle swarm op-
Storage, Vol.26, p.100951 (2019) [Link] timization”, J. Power Electron. (2017) [Link]
( 26 ) I. Jorge, A. Samet, T. Mesbahi, and R. Bone: “New ANN results on a ma- 5.1288
jor benchmark for the prediction of RUL of Lithium Ion batteries in electric ( 46 ) B. Saha, K. Goebel, and J. Christophersen: “Comparison of prognostic algo-
vehicles”, Proc. - 19th IEEE Int. Conf. Mach. Learn. Appl. ICMLA 2020, rithms for estimating remaining useful life of batteries”, Trans. Inst. Meas.
pp.1246–1253 (2020) [Link] Control (2009) [Link]
( 27 ) S. Shen, X. Chen, C. Hu, M.K. Sadoughi, and M. Hong: “Online estimation ( 47 ) X. Zheng and H. Fang: “An integrated unscented kalman filter and relevance
of lithium-ion battery capacity using deep convolutional neural networks”, In: vector regression approach for lithium-ion battery remaining useful life and
Proceedings of the ASME Design Engineering Technical Conference (2018) short-term capacity prediction”, Reliab. Eng. Syst. Saf., Vol.144, pp.74–82
( 28 ) J. Wu, C. Zhang, and Z. Chen: “An online method for lithium-ion battery (2015) [Link]
remaining useful life estimation using importance sampling and neural net- ( 48 ) J. Xu, A. Zhen, Z. Cai, P. Wang, K. Gao, and D. Jiang: “State of Health
works”, Appl. Energy, Vol.173, pp.134–140 (2016) [Link] Diagnosis and Remaining Useful Life Prediction of Lithium-Ion Batteries
[Link].2016.04.057 Based on Multi-Feature Data and Mechanism Fusion”, IEEE Access., Vol.9,
( 29 ) J. Liu, A. Saxena, K. Goebel, B. Saha, and W. Wang: “An adaptive recurrent pp.85431–85441 (2021) [Link]
neural network for remaining useful life prediction of lithium-ion batteries”, ( 49 ) B. Gou, Y. Xu, S. Fang, R.A. Pratama, and S. Liu: “Remaining Useful
Annu. Conf. Progn. Heal. Manag. Soc. PHM 2010, p.9 (2010) Life Prediction for Lithium-ion Battery Using Ensemble Learning Method”,
( 30 ) D. Chen, W. Hong, and X. Zhou: “Transformer Network for Remain- IEEE Power Energy Soc. Gen. Meet. (2019) [Link]
ing Useful Life Prediction of Lithium-Ion Batteries”, IEEE Access, Vol.10, PESGM40551.2019.8973811
pp.19621–19628 (2022) [Link] ( 50 ) B. Gou, Y. Xu, and X. Feng: “State-of-Health Estimation and Remaining-
( 31 ) J. Zhou, D. Liu, Y. Peng, and X. Peng: “Dynamic battery remaining useful Useful-Life Prediction for Lithium-Ion Battery Using a Hybrid Data-Driven
life estimation: An on-line data-driven approach”, 2012 IEEE I2MTC - Int. Method”, IEEE Trans. Veh. Technol., Vol.69, pp.10854–10867 (2020)
Instrum. Meas. Technol. Conf. Proc., pp.2196–2199 (2012) [Link] [Link]
10.1109/I2MTC.2012.6229280 ( 51 ) T. Sun, B. Xia, Y. Liu, Y. Lai, W. Zheng, H. Wang, W. Wang, and M. Wang:
( 32 ) A. Nuhic, T. Terzimehic, T. Soczka-Guth, M. Buchholz, and K. Dietmayer: “A novel hybrid prognostic approach for remaining useful life estimation
“Health diagnosis and remaining useful life prognostics of lithium-ion bat- of lithium-ion batteries”, Energies, Vol.12 (2019) [Link]
teries using data-driven methods”, J. Power Sources, Vol.239, pp.680–688 en12193678
(2013) [Link] ( 52 ) Q. Xue, S. Shen, G. Li, Y. Zhang, Z. Chen, and Y. Liu: “Remaining Useful
( 33 ) M.A. Patil, P. Tagade, K.S. Hariharan, S.M. Kolake, T. Song, T. Yeo, and S. Life Prediction for Lithium-Ion Batteries Based on Capacity Estimation and
Doo: “A novel multistage Support Vector Machine based approach for Li ion Box-Cox Transformation”, IEEE Trans. Veh. Technol., Vol.69, pp.14765–
battery remaining useful life estimation”, Appl. Energy (2015) [Link] 14779 (2020) [Link]
org/10.1016/[Link].2015.08.119 ( 53 ) Z. Xue, Y. Zhang, C. Cheng, and G. Ma: “Remaining useful life prediction
( 34 ) Y. Li, S. Wang, W. Chen, and W. Han: “Holistic Comparison of Different of lithium-ion batteries with adaptive unscented kalman filter and optimized
Kernel Functions for Support Vector Regression Based on State-of-health support vector regression”, Neurocomputing (2020) [Link]
Prediction of Lithium-ion Battery”, Proc. - 11th Int. Conf. Progn. Syst. Heal. [Link].2019.09.074
Manag. PHM-Jinan 2020, pp.40–46 (2020) [Link] ( 54 ) Z. Wang, S. Zeng, J. Guo, and T. Qin: “Remaining capacity estimation of
Jinan48558.2020.00015 lithium-ion batteries based on the constant voltage charging profile”, PLoS
( 35 ) S. Abdelghafar, E. Goda, A. Darwish, and A.E. Hassanien: “Satellite One (2018) [Link]
Lithium-ion Battery Remaining Useful Life Estimation by Coyote Opti- ( 55 ) M.E. Tipping: “Sparse Bayesian Learning and the Relevance Vector Ma-
mization Algorithm”, Proc. - 2019 IEEE 9th Int. Conf. Intell. Comput. Inf. chine”, J. Mach. Learn. Res. (2001)
Syst. ICICIS 2019, pp.124–129 (2019) [Link] ( 56 ) L. Zhao, Y. Wang, and J. Cheng: “A hybrid method for remaining useful life
2019.9014752 estimation of lithium-ion battery with regeneration phenomena”, Appl. Sci.
( 36 ) Y. Wang, Y. Ni, S. Lu, J. Wang, and X. Zhang: “Remaining Useful Life Pre- (2019) [Link]
diction of Lithium-Ion Batteries Using Support Vector Regression Optimized ( 57 ) P. Guo, Z. Cheng, and L. Yang: “A data-driven remaining capacity estimation
by Artificial Bee Colony”, IEEE Trans. Veh. Technol. (2019) [Link] approach for lithium-ion batteries based on charging health feature extrac-
10.1109/TVT.2019.2932605 tion”, J. Power Sources (2019) [Link]
( 37 ) J. Du, W. Zhang, C. Zhang, and X. Zhou: “Battery remaining useful life pre- 072
diction under coupling stress based on support vector regression”, In: Energy ( 58 ) Y. Liu, G. Zhao, and X. Peng: “Deep learning prognostics for lithium-ion bat-
Procedia (2018) tery based on ensembled long short-term memory networks”, IEEE Access,
( 38 ) Y.Z. Zhang, R. Xiong, H.W. He, and M. Pecht: “Validation and verification of Vol.7, pp.155130–155142 (2019) [Link]
a hybrid method for remaining useful life prediction of lithium-ion batteries”, 2937798
J. Clean. Prod. (2019) [Link] ( 59 ) B. Chinomona, C. Chung, L.K. Chang, W.C. Su, and M.C. Tsai: “Long short-
( 39 ) L. Chen, Y. Zhang, Y. Zheng, X. Li, and X. Zheng: “Remaining useful life term memory approach to estimate battery remaining useful life using par-
prediction of lithium-ion battery with optimal input sequence selection and tial data”, IEEE Access, Vol.8, pp.165419–165431 (2020) [Link]
error compensation”, Neurocomputing, Vol.414, pp.245–254 (2020) https:// 1109/ACCESS.2020.3022505
[Link]/10.1016/[Link].2020.07.081 ( 60 ) J. Qiao, X. Liu, and Z. Chen: “Prediction of the Remaining Useful
( 40 ) H. Chaoui, N. Golbon, I. Hmouz, R. Souissi, and S. Tahar: “Lyapunov-based Life of Lithium-Ion Batteries Based on Empirical Mode Decomposition
adaptive state of charge and state of health estimation for lithium-ion bat- and Deep Neural Networks”, IEEE Access, Vol.8, pp.42760–42767 (2020)
teries”, IEEE Trans. Ind. Electron. (2015) [Link] [Link]
2341576 ( 61 ) X. Song, F. Yang, D. Wang, and K.L. Tsui: “Combined CNN-LSTM Net-
( 41 ) D. Liu, J. Zhou, D. Pan, Y. Peng, and X. Peng: “Lithium-ion battery re- work for State-of-Charge Estimation of Lithium-Ion Batteries”, IEEE Access
maining useful life estimation with an optimized Relevance Vector Machine (2019) [Link]
algorithm with incremental learning”, Meas. J. Int. Meas. Confed., Vol.63, ( 62 ) L. Li, Y. Peng, Y. Song, and D. Liu: “Lithium-Ion Battery Remaining Use-
pp.143–151 (2015) [Link] ful Life Prognostics Using Data-Driven Deep Learning Algorithm”, Proc. -

573 IEEJ Journal IA, Vol.12, No.4, 2023

 Machine Learning-based Remaining Useful Life Estimation Akash Samanta et al.

2018 Progn. Syst. Heal. Manag. Conf. PHM-Chongqing 2018, pp.1094–1100 ( 77 ) J. Zhu, T. Tan, L. Wu, and H. Yuan: “RUL prediction of lithium-ion bat-
(2019) [Link] tery based on improved DGWO-ELM method in a random discharge rates
( 63 ) D. Zhou, Z. Li, J. Zhu, H. Zhang, and L. Hou: “State of Health Monitor- environment”, IEEE Access (2019) [Link]
ing and Remaining Useful Life Prediction of Lithium-Ion Batteries Based 2936822
on Temporal Convolutional Network”, IEEE Access, Vol.8, pp.53307–53320 ( 78 ) L. Chen, H. Wang, J. Chen, J. An, B. Ji, Z. Lyu, W. Cao, and H. Pan: “A
(2020) [Link] novel remaining useful life prediction framework for lithium-ion battery us-
( 64 ) P. Li, Z. Zhang, Q. Xiong, B. Ding, J. Hou, D. Luo, Y. Rong, and S. ing grey model and particle filtering”, Int. J. Energy Res., Vol.44, pp.7435–
Li: “State-of-health estimation and remaining useful life prediction for 7449 (2020) [Link]
the lithium-ion battery based on a variant long short term memory neu- ( 79 ) Y. Li, C. Zou, M. Berecibar, E. Nanini-Maury, J.C.W. Chan, P. van den
ral network”, J. Power Sources, Vol.459, p.228069 (2020) [Link] Bossche, J. Van Mierlo, and N. Omar: “Random forest regression for
10.1016/[Link].2020.228069 online capacity estimation of lithium-ion batteries”, Appl. Energy (2018)
( 65 ) J. Qu, F. Liu, Y. Ma, and J. Fan: “A Neural-Network-Based Method for [Link]
RUL Prediction and SOH Monitoring of Lithium-Ion Battery”, IEEE Ac- ( 80 ) E. Martinez-Laserna, I. Gandiaga, E. Sarasketa-Zabala, J. Badeda, D.I. Stroe,
cess, Vol.7, pp.87178–87191 (2019) [Link] M. Swierczynski, and A. Goikoetxea: “Battery second life: Hype, hope or re-
2925468 ality?”, A critical review of the state of the art (2018)
( 66 ) G. Ma, Y. Zhang, C. Cheng, B. Zhou, P. Hu, and Y. Yuan: “Remaining useful ( 81 ) I. Sanz-Gorrachategui, P. Pastor-Flores, M. Pajovic, Y. Wang, P.V. Orlik, C.
life prediction of lithium-ion batteries based on false nearest neighbors and Bernal-Ruiz, A. Bono-Nuez, J.S. Artal-Sevil: “Remaining Useful Life Es-
a hybrid neural network”, Appl. Energy., Vol.253, p.113626 (2019) https:// timation for LFP Cells in Second-Life Applications”, IEEE Trans. Instrum.
[Link]/10.1016/[Link].2019.113626 Meas., Vol.70 (2021) [Link]
( 67 ) L. Ren, L. Zhao, S. Hong, S. Zhao, H. Wang, and L. Zhang: “Remain-
ing Useful Life Prediction for Lithium-Ion Battery: A Deep Learning
Approach”, IEEE Access., Vol.6, pp.50587–50598 (2018) [Link]
10.1109/ACCESS.2018.2858856 Akash Samanta (Non-member) (Student Member, IEEE) received
( 68 ) Z. Huang, X. Zhou, D. Gao, X. Zhang, F. Jiang, B. Chen, Y. Yang,
[Link]. degree in Electrical Engineering from West
M. Wu, and J. Peng: “A Novel Label-Free Supervision Learning Method
Bengal University of Technology in 2012 and MBA
for Lithium-ion Battery RUL Prediction”, 2019 IEEE Energy Convers.
Congr. Expo. ECCE 2019, pp.3150–3156 (2019) [Link]
degree in Energy Management and [Link]. degree
ECCE.2019.8912697 in Electrical Engineering and from University of Cal-
( 69 ) X. Li, L. Zhang, Z. Wang, and P. Dong: “Remaining useful life prediction for cutta in 2014 and 2018 respectively. Formerly, he was
lithium-ion batteries based on a hybrid model combining the long short-term appointed as Project Officer and Solar Energy Mas-
memory and Elman neural networks”, J. Energy Storage, Vol.21, pp.510–518 ter Trainer in the Department of Energy Management,
(2019) [Link] Indian Institute of Social Welfare and Business Man-
( 70 ) J. Hong, D. Lee, E.R. Jeong, and Y. Yi: “Towards the swift prediction of the agement, Kolkata, India. He is currently a Doctoral
remaining useful life of lithium-ion batteries with end-to-end deep learning”, Research Scholar with the Department of Electrical, Computer and Software
Appl. Energy, Vol.278, p.115646 (2020) [Link] Engineering, Ontario Tech University, Oshawa, ON, Canada. His research
2020.115646 interest includes, electric energy storage systems, battery management sys-
( 71 ) P. Khumprom and N. Yodo: “A data-driven predictive prognostic model for tem, power electronics converters, and the application of machine learning
lithium-ion batteries based on a deep learning algorithm”, Energies (2019)
and artificial intelligence in battery management system and power electron-
[Link]
ics converters.
( 72 ) K. Park, Y. Choi, W.J. Choi, H.Y. Ryu, and H. Kim: “LSTM-Based Bat-
tery Remaining Useful Life Prediction with Multi-Channel Charging Pro-
files”, IEEE Access, Vol.8, pp.20786–20798 (2020) [Link]
ACCESS.2020.2968939 Sheldon Williamson (Non-member) (Fellow, IEEE) received the B.E.
( 73 ) Y. Che, Z. Deng, X. Lin, L. Hu, and X. Hu: “Predictive Battery Health Man- degree (Hons.) in electrical engineering from the
agement with Transfer Learning and Online Model Correction”, IEEE Trans. University of Mumbai, Mumbai, India, in 1999, and
Veh. Technol. (2021) [Link] the M.S. and Ph.D. degrees (Hons.) in electrical en-
( 74 ) Y. Zhang, R. Xiong, H. He, and M.G. Pecht: “Long short-term memory gineering from the Illinois Institute of Technology,
recurrent neural network for remaining useful life prediction of lithium- Chicago, IL, USA, in 2002 and 2006, respectively.
ion batteries”, IEEE Trans. Veh. Technol., Vol.67, pp.5695–5705 (2018)
He is currently a Professor with the Department of
[Link]
Electrical, Computer and Software Engineering and
( 75 ) Y. Cao, B. Zhao, L. Wang, H. Shen, and J. Hu: “Research on Battery Life
prediction Based on Deep Learning”, Asia-Pacific Power Energy Eng. Conf.
the Director of Smart Transportation Electrification
APPEEC. 2020-Septe (2020) [Link] and Energy Research (STEER) Group, Faculty of En-
9220734 gineering and Applied Sciences, Ontario Tech University, Oshawa, ON,
( 76 ) J. Fan, J. Fan, F. Liu, J. Qu, and R. Li: “A Novel Machine Learning Method Canada. His current research interests include advanced power electronics,
Based Approach for Li-Ion Battery Prognostic and Health Management”, electric energy storage systems, and motor drives for transportation electrifi-
IEEE Access, Vol.7, pp.160043–160061 (2019) [Link] cation. He holds the prestigious NSERC Canada Research Chair position in
ACCESS.2019.2947843 electric energy storage systems for transportation electrification.

574 IEEJ Journal IA, Vol.12, No.4, 2023